SIP trunking connects a company's phone system to the public telephone network over an IP connection, using the Session Initiation Protocol (SIP). It replaces physical phone lines with virtual channels, and it is also a common way to connect an AI voice agent to a contact center.
A "trunk" is the shared connection that carries many calls at once between two systems. Its capacity is usually counted in concurrent calls, often called channels.
How SIP trunking works
Signaling. SIP sets up, changes, and ends calls. RFC 3261 (June 2002) defines it as an application-layer signaling protocol. The voice itself travels separately, usually as RTP packets, and often on a different path from the signaling.
The trunk. The company's PBX or contact center platform registers with a carrier, or peers with it at fixed addresses. The SIP Forum's SIPconnect recommendation, version 2.0 ratified in December 2016, sets out how IP PBXs and service provider networks connect. It covers codecs, keypad tones, call transfer, and security.
The session border controller (SBC). Most enterprise trunks end at an SBC, which sits at the border between networks. RFC 5853 (April 2010) describes the functions SBCs commonly provide. They handle signaling and media at the edge, enforce policy, and often hide the internal network from the outside.
How an AI voice agent connects over SIP
The agent behaves like one more SIP destination on the network. A typical call arrives from the carrier at the SBC, as it does today. The contact center platform or conversational IVR answers, picks the call reason, and sends selected calls to the agent over a SIP trunk.
To hand the call back, the agent can send a SIP REFER. RFC 3515 (April 2003) defines REFER and names call transfer as one of its uses. The platform then routes the call to a queue or an extension. The other option is a bridged transfer, where the agent's platform stays in the call.
Context for the human agent can travel in SIP headers. RFC 7433 (January 2015) defines a User-to-User header for this, and allows proxies and other intermediaries to remove it. A sturdier pattern sends a short session ID with the call and lets the platform fetch the rest through an API (see intelligent handoff).
What to check before connecting a voice agent
Codec. G.711, ITU-T's pulse code modulation of voice frequencies, is widely supported. Transcoding between codecs on the way to the agent can add delay and lower audio quality.
Keypad tones. Callers still type PINs and account numbers. RFC 4733 (December 2006, replacing RFC 2833) carries keypad tones as RTP events. It helps to confirm both sides use the same method.
Transfer method. Some trunks and SBCs block REFER by policy. It is worth deciding between REFER and bridged transfers before the first test call.
Security. If the SBC requires SIP over TLS and encrypted media (SRTP), both sides need matching versions, certificates, and key exchange. A mismatch often shows up as a call with no audio.
Capacity and recording. Channels should cover peak hours, such as an outage spike. It also helps to agree where the AI part of the call is recorded, and for how long.
Frequently asked questions
What is the difference between SIP and VoIP?
VoIP is the broad term for voice calls carried over IP networks. SIP is one protocol used to set up, manage, and end those calls. SIP trunking is the service that links a company's phone system to a carrier using SIP.
Do we need a new SIP trunk to use an AI voice agent?
Often not. In many setups, the existing contact center platform or SBC adds the agent as a new SIP destination and routes chosen calls to it. The work is mostly configuration: addresses, codecs, transfer rules, and how context is passed.
What is a session border controller?
A session border controller (SBC) is a device or software that sits where a company's voice network meets a carrier or another network. It handles SIP signaling and media, enforces security and routing policy, and can convert between codecs.
Can an AI voice agent transfer a call to a human agent over SIP?
Yes. The usual methods are a SIP REFER, which asks the platform to make the transfer, or a bridged transfer. It helps to test that the call reason and caller details reach the human agent through the real path, SBC included.
How we approach this at Zowie
Zowie's AI agents work on voice, chat, and email, with over 100 million conversations a year across our clients. On a call, the language model speaks. A separate, deterministic Decision Engine returns every amount, rate, offer, disclosure, and escalation from your rules, with an audit log. To talk through your telephony setup, you are welcome to book a demo.
Explore: Conversational IVR, Call routing, Intelligent handoff, Turn-taking, Contact center AI, Voice AI